Add UnifiedBuffer!T for CUDA Unified Memory (Managed Memory)

[badnikhil]

Summary

Adds UnifiedBuffer!T to the CUDA driver, enabling allocation of Unified Memory (Managed Memory) via cuMemAllocManaged. A unified buffer is accessible from both the host (CPU) and device (GPU) through the same pointer, so explicit copy!(Copy.hostToDevice) / copy!(Copy.deviceToHost) calls are not needed.

Requires CUDA Compute Capability >= 3.0 (Device.supportsUnifiedMemory == true).

---

tests/main.d

• Added a unified memory correctness test block, guarded by a runtime dev.supportsUnifiedMemory check
• Runs saxpy kernel via UnifiedBuffer, verifies result via hostSlice with no explicit copy calls, synchronized with Context.sync()

---

Notes

• prefetch() is a stub with a doc comment explaining the limitation.
• A follow-up to migrate from derelict-cuda to a bindbc-based CUDA binding is being tracked separately.

[badnikhil]

i tried to document everything precisely in comments. still i would request you to read everything and provide feedback!